RSAFormer: A method of polyp segmentation with region self-attention transformer

• Published in: Computers in biology and medicine Vol. 172; p. 108268
• Main Authors: Yin, Xuehui, Zeng, Jun, Hou, Tianxiao, Tang, Chao, Gan, Chenquan, Jain, Deepak Kumar, García, Salvador
• Format: Journal Article
• Language: English
• Published: United States Elsevier Ltd 01.04.2024; Elsevier Limited
• Subjects: Colon cancer; Colonoscopy; Feature extraction; Feature maps; Image Processing, Computer-Assisted; Information processing; Modules; Polyp segmentation; Polyps; Region self-attention; Segmentation; Transformer; Transformers; Uncertainty; Region self-attention; Transformer; Colonoscopy; Polyp segmentation
• ISSN: 0010-4825; 1879-0534
• DOI: 10.1016/j.compbiomed.2024.108268

Colonoscopy has attached great importance to early screening and clinical diagnosis of colon cancer. It remains a challenging task to achieve fine segmentation of polyps. However, existing State-of-the-art models still have limited segmentation ability due to the lack of clear and highly similar boundaries between normal tissue and polyps. To deal with this problem, we propose a region self-attention enhancement network (RSAFormer) with a transformer encoder to capture more robust features. Different from other excellent methods, RSAFormer uniquely employs a dual decoder structure to generate various feature maps. Contrasting with traditional methods that typically employ a single decoder, it offers more flexibility and detail in feature extraction. RSAFormer also introduces a region self-attention enhancement module (RSA) to acquire more accurate feature information and foster a stronger interplay between low-level and high-level features. This module enhances uncertain areas to extract more precise boundary information, these areas being signified by regional context. Extensive experiments were conducted on five prevalent polyp datasets to demonstrate RSAFormer’s proficiency. It achieves 92.2% and 83.5% mean Dice on Kvasir and ETIS, respectively, which outperformed most of the state-of-the-art models. •A region self-attention enhancement network (RSAFormer) is proposed.•RSAFormer uniquely uses a dual decoder structure to generate various feature maps.•A novel region self-attention enhancement (RSA) module is proposed.